Intel LLM-Scaler-vLLM PV 1.5 updates vLLM and PyTorch versions
1 min read
•
Impact: 5/10
Key Takeaway
Intel's containerized stack for Arc Graphics significantly updates its core vLLM and PyTorch components while adding new model support and quantization methods.
Summary
Intel software engineers released LLM-Scaler-vLLM PV 1.5, a Docker-based stack designed for running Large Language Models on Intel Arc Graphics hardware. This update advances the stack closer to upstream by jumping from vLLM 0.14 to 0.26 and PyTorch 2.10 to 2.12. The release also introduces support for specific Qwen and Gemma models, along with new quantization capabilities.
Technical improvements in PV 1.5
- Upgraded vLLM from 0.14 to 0.26 and PyTorch from 2.10 to 2.12.
- Added support for Qwen3.6-27B, Qwen3.8-27B, Qwen3.6-35B-A3B, gemma-4-26B-A4B-it, and gemma-4-31B-it.
- Introduced support for online FP8 and symmetric INT4 quantization.
- Added FP8 KV cache support.
- Improved performance for time-to-first-token (Qwen/Gemma 4), MTP decoding, LoRA serving, and FP8 KV cache.
Why it matters
The update brings Intel Arc users closer to the upstream vLLM ecosystem and improves the efficiency of running LLMs on Intel consumer graphics hardware.